随着数字化水平的提高，文本数据呈现出爆炸式增长态势。在国家政策、行业资本加速推进挖掘规模化数据价值的背景下，文本信息抽取技术的重要性正在显著提高。虽然基于神经网络模型开发的方法已经得到了广泛的应用，但仍然面临着两个重要挑战：其一，处理具有复杂层次结构的文本时性能不佳；其二，信息抽取过程缺乏可解释性，抽取结果的信息表达能力有限。受到认知领域中关于自然语言处理的发现启发，本文以层次化编码和神经符号化处理技术为核心，系统性地开展了以下研究工作：受到大脑皮层对自然语言短语的响应特点启发，针对嵌套命名实体识别问题中复杂重叠关系带来的计算复杂，准确率低的问题，本文提出了局部超图构建网络模型，利用局部超图抓取短语结构特点进行编码，通过预测动作符号序列控制超图生成，完成识别任务。实验结果证明，本文提出的模型在嵌套命名实体识别任务中取得了最优性能，并降低了识别过程的时间复杂度。受到人在阅读过程中采用精读与略读组合的文本处理策略启发，针对实际应用场景中需要对目标片段进行抽取的问题，本文提出了变焦神经网络模型，利用文本的篇章结构进行层次化编码，通过预测多个尺度的动作符号灵活地进行序列标注，完成识别。实验结果证明，本文提出的模型在目标片段识别任务中优于序列标注基线模型。同时，经过强化学习方法的训练，模型可以有效模拟人的阅读过程。受到认知领域中有关事件认知的系统性理论启发，基于文本事件抽取，本文提出了与应用场景结合更加紧密，更加复杂的文本语义解析问题。针对这一问题，本文提出了面向对象的神经编码解析器模型，利用与事件表征相对应的图结构进行属性、对象等多个层次的编码，通过预测动作符号顺序完成对符号化记忆、向量化记忆的更新和事件表征结构的生成，最终完成对事件信息的抽取。实验证明，本文提出的模型可以有效识别复杂的事件信息，性能显著优于没有针对性结构设计的基线模型。综上所述，在认知领域发现的启发下，本文提出了多种神经符号化模型，在提高信息抽取方法的性能、可解释性和表达能力上做出了一定的贡献。

With the improvement of digitalization, textual data has shown an explosive growth trend. Against the backdrop of national policies and industry capital accelerating the mining of large-scale data value, the importance of text information extraction technology is significantly increasing. Although methods based on neural network models have been widely applied, they still face two important challenges: (1) They show poor performance when dealing with text with complex hierarchical structures; (2) They lack interpretability in the information extraction process and the information expression ability of their extraction results is limited. Inspired by the findings in the cognitive field of natural language processing, this paper systematically carries out the following research works with hierarchical encoding and neural symbolic processing techniques as the core: Inspired by the response characteristics of the brain cortex to natural language phrases, this paper proposes Local Hypergraph Building Network(LHBN) to address the problem of high computational complexity and low accuracy caused by complex overlapping relationships in nested named entity recognition. The model leverages local hypergraphs to capture the characteristics of phrase structure for encoding, and generates hypergraphs by predicting action symbol sequences to complete the recognition task. Experimental results demonstrate that the proposed model achieves state-of-the-art performance in nested named entity recognition tasks and reduces the time complexity of the recognition process. Inspired by the text processing strategy that humans use a combination of intensive reading and skimming during the reading process, this paper proposes Zooming Neural Network(ZoomNet) to address the problem of extracting target fragments in practical application scenarios. The model uses text structure for hierarchical encoding and flexibly performs sequence labeling by predicting action symbols at multiple scales to complete the recognition task. Experimental results demonstrate that the proposed model outperforms the sequence labeling baseline model in target fragment recognition tasks. Moreover, after training with reinforcement learning methods, the model can effectively simulate the reading process of humans. Inspired by the systematic theories on event cognition in the cognitive field, this paper proposes a text parsing task with tighter integration with application scenarios and stronger information expression capability based on text event extraction. To address this problem, this paper proposes Object-oriented Neural Programming Parser, which uses a graph structure corresponding to event representation for encoding at multiple levels such as attributes and objects. The model completes event information extraction by predicting the sequence of action symbols for updating symbolic memory, vectorized memory, and generating event representation structures. Experimental results demonstrate that the proposed model can effectively recognize complex event information and outperforms baseline models without targeted structure design.In conclusion, inspired by the findings in the cognitive field, this paper proposes various neural symbolic models that contribute to improving the performance, interpretability, and expressiveness of information extraction methods.